AIDE2 memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

All positive findings are claimed. The accepted members retain doctrine/design and author-reported operation from the frozen paper, not inspected implementation, observed raw runs or causal experiments; see RT-RTE-harness-admission and MEM-RTE-summary-delivery. Their operation reports do not justify upgrading a profile basis to observed. Coverage assessments describe classification completeness within each named unit, not execution reliability.

The inventory separates reduced history, failure fragments, reviewer text/scores, selector metadata, outer history and retained executable organization. A codebase or assembled prompt does not identify its physical substrate. All storage units therefore remain not-determinable. Fixed provider parameters do not enter this use-derived memory boundary; the profile does not claim an absence of provider state.

The reconciliation supersedes the epistemic summary and signature declarations with MEM-OBJ-candidate-summary and MEM-OBJ-error-signatures. Findings cite those canonical records. Reviewer feedback and evaluation scores remain overlapping material parts rather than whole-record duplicates. MEM-OBJ-review-feedback distinguishes numerical selection from interpreted feedback; EPI-OBJ-evaluation-scores also covers private grade aggregation.

Consolidation classifies selection/compression of existing history, not proven semantic faithfulness. Unknown summary authorship remains beside automatic context admission. The signature route separately establishes extraction, deduplication, admission and later consumption. Neither mechanism establishes synthesis, retained causal explanations, expiry or deletion. The summary chain's trace-learning classification remains not-determinable because automatic later delivery does not establish automatic summary production. Its trace_source classification likewise remains unresolved: recorded execution input alone cannot establish a source value for a write that has not been shown to qualify. Consolidation, trace-extracted lineage and later knowledge consumption remain supported independently. See MEM-RTE-summary-delivery and MEM-RTE-failure-injection.

Read-back direction separates unsolicited harness supply from unspecified artifact exploration. Root/recency selection and threshold/recurrence filtering are coarse push signals. Text equality used to deduplicate errors is not a lexical retrieval query; node labels do not establish identifier-based delivery. Unknown exploration and alternative role prompts prevent a complete push-only conclusion.

Outer learning is a different route: evaluation experience guides automatic code rewriting and grade-controlled adoption, and accepted code governs subsequent work. Updating the incumbent is evolution; selecting a proposal as incumbent is promotion. Prior code/grade sequences support a trajectories finding, while execution/scoring feedback supports tool-traces. Opaque evaluation artifacts prevent a complete source inventory. The relevant route is RT-RTE-harness-admission, with downstream force in RT-BAP-incumbent-code.

The robustness reminder is included because it is embedded in retained, automatically admitted successor code, not because static shipped instructions alone count as accumulated memory. Its standing instruction force differs from derived history's knowledge force. The short-code criterion concerns generated task scripts, not outer successor admission; its validation and enforced retry are separate from private-grade selection. See RT-RTE-robustness-controls.

No axis claims bounded absence. Unknown initialization, summary fields/control, artifact payloads, alternative-agent prompts, feedback delivery, archive retention and reload prevent stronger inventory conclusions. Positive trace-learning findings establish the documented trace-fed write/consumer routes, not independently verified improved capacity, component benefit or weight updates.

Source-read log: none. Classification used only supplied accepted members and their reconciliation; no capture read, new evidence or new record was needed.